Responsible Intelligence in Practice: A Fairness Audit of Open Large Language Models for Library Reference Services

1. Problem Statement

As libraries increasingly adopt Large Language Models (LLMs) to scale reference services, a critical ethical question arises: Do these AI systems provide equitable service to all patrons regardless of demographic identity?

• Context: Historical audit studies of human librarians have documented disparities in service quality based on perceived race and gender (e.g., African American or Arab-sounding names receiving less complete or courteous responses).
• Risk: LLMs are trained on vast corpora containing societal biases, which they may reproduce, potentially undermining library commitments to impartial access.
• Gap: While general AI fairness benchmarks exist, there is a lack of domain-specific evaluation for library reference services, particularly regarding open-source LLMs that libraries might deploy on-premises to maintain data control and governance.

2. Methodology

The authors employed a two-stage Fairness Evaluation Protocol (FEP) to audit three open LLMs across academic and public library contexts.

A. Experimental Design & Data Synthesis

• Models Evaluated: Three state-of-the-art open LLMs capable of local deployment:
  1. Llama-3.1 8B (Meta)
  2. Gemma-2 9B (Google)
  3. Ministral 8B (Mistral AI)
• Simulation Setup:
  • Query Templates: Synthesized 6 types of reference emails (3 academic, 3 public) covering topics like collection research, sports history, population data, and digital literacy (e.g., resume uploading, password recovery).
  • Demographic Cues: User names were algorithmically synthesized to signal Sex (Male/Female) and Race/Ethnicity (6 groups: White, Black/African American, Asian/PI, American Indian/Alaska Native, Two+ Races, Hispanic/Latino).
  • Cohort: 12 demographic groups (2 sexes × 6 races/ethnicities) with balanced sampling using SSA baby name data and US Census surname distributions.
  • Prompting: Models were instructed via system prompts to act as "helpful, respectful, and honest librarians" from specific institutions.
• Corpus Generation: 2,500 responses per model per setting (Academic/Public), generated with 5 random seeds and a temperature of 0.7.

B. Evaluation Protocol (FEP)

The protocol consists of two stages:

1. Stage 1: Diagnostic Classification (Detection)

   • Goal: Determine if demographic attributes can be predicted from the model's response text better than chance.
   • Features: TF-IDF vectors of the top 120 content words (top 60 for Gemma-2 due to shorter length). Gendered honorifics were masked to prevent trivial shortcuts.
   • Classifiers: Three models were trained to predict the user's demographic group: Logistic Regression (LogReg), Multi-Layer Perceptron (MLP), and XGBoost.
   • Metric: Classification accuracy compared against chance levels (50% for binary sex; ~16.7% for 6-class race).
   • Significance: Statistical significance tested via t-tests with Bonferroni correction ($\alpha = 0.0056$).

2. Stage 2: Interpretation (Attribution)

   • Goal: If Stage 1 detects significant bias, identify the specific linguistic markers driving the difference.
   • Method: Fitting a statistical logistic regression model to compute coefficients and p-values for individual words.
   • Visualization: Volcano plots were used to identify words that are both statistically significant and have a meaningful effect size (odds ratio $\ge$ 2 or $\le$ 0.5).

3. Key Results

A. Racial and Ethnic Fairness

• Finding: No systematic differentiation was detected across any of the three models in either academic or public library settings.
• Data: Classification margins for race/ethnicity hovered near zero for all classifiers (LogReg, MLP, XGBoost). None of the 18 tests (3 models × 3 classifiers × 2 settings) reached statistical significance after correction.
• Conclusion: The models do not appear to vary their responses based on racial or ethnic cues embedded in names.

B. Sex-Based Fairness

• Finding: Generally equitable, with one minor exception.
  • Ministral 8B & Gemma-2: Showed no detectable sex-linked differentiation in either setting.
  • Llama-3.1 (Academic Setting): Showed statistically significant differentiation (LogReg: +3.64%, XGBoost: +5.76% above chance).
• Root Cause Analysis:
  • The differentiation was driven almost exclusively by the salutation "dear."
  • Pattern: "Dear" appeared in 66.2% of responses to female users vs. 48.4% for male users in the academic setting.
  • Nature of Bias: This is a surface-level stylistic variation (politeness strategy) rather than a difference in the substantive quality, completeness, or helpfulness of the reference advice.
  • Context Dependence: This pattern did not appear in the public library setting, suggesting the bias is context-specific to the academic framing.

4. Key Contributions

1. Domain-Specific Audit: Provides the first systematic fairness audit of open LLMs specifically tailored to the library reference service domain, bridging the gap between general AI benchmarks and professional library ethics.
2. Methodological Framework: Demonstrates the utility of the Fairness Evaluation Protocol (FEP), combining diagnostic classification with interpretive linguistic analysis to distinguish between harmful bias and benign contextual variation.
3. Empirical Evidence on Open Models: Offers evidence that contemporary open LLMs (Llama-3.1, Gemma-2, Ministral) largely avoid the racial/ethnic biases historically found in human reference interactions, though minor stylistic sex biases may persist.
4. Governance Guidance: Argues that "responsible intelligence" in libraries requires ongoing monitoring rather than one-time certification, emphasizing that open models allow institutions to retain control for local auditing.

5. Significance and Implications

• Professional Trust: The findings suggest that open LLMs can be a viable tool for scaling library services without immediately replicating the demographic disparities found in human-led services, provided they are monitored.
• Nuanced Understanding of Bias: The study highlights that not all statistical differences constitute "harmful discrimination." The detected sex bias in Llama-3.1 was a stylistic artifact (politeness) rather than a denial of service, illustrating the need for human librarians to interpret audit results.
• Limitations & Future Work:
  • Context Specificity: The study used standard English and name-based cues. Bias may emerge with dialects (e.g., African American English) or explicit identity markers, which were not tested.
  • Intersectionality: The study treated sex and race as separate dimensions; future work must examine intersectional effects (e.g., Black women vs. White men).
  • Dynamic Nature: Fairness is not static; as models evolve and training data shifts, libraries must commit to continuous auditing.
• Call to Action: Libraries should adopt a participatory governance model, combining computational audits with professional judgment and community feedback to ensure AI tools advance, rather than undermine, the mission of equitable information access.